Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2022 The HuggingFace Inc. team. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15"""BiT model configuration"""16 17from ...configuration_utils import PretrainedConfig18from ...utils import logging19from ...utils.backbone_utils import BackboneConfigMixin, get_aligned_output_features_output_indices20 21 22logger = logging.get_logger(__name__)23 24 25class BitConfig(BackboneConfigMixin, PretrainedConfig):26    r"""27    This is the configuration class to store the configuration of a [`BitModel`]. It is used to instantiate an BiT28    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the29    defaults will yield a similar configuration to that of the BiT30    [google/bit-50](https://huggingface.co/google/bit-50) architecture.31 32    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the33    documentation from [`PretrainedConfig`] for more information.34 35    Args:36        num_channels (`int`, *optional*, defaults to 3):37            The number of input channels.38        embedding_size (`int`, *optional*, defaults to 64):39            Dimensionality (hidden size) for the embedding layer.40        hidden_sizes (`list[int]`, *optional*, defaults to `[256, 512, 1024, 2048]`):41            Dimensionality (hidden size) at each stage.42        depths (`list[int]`, *optional*, defaults to `[3, 4, 6, 3]`):43            Depth (number of layers) for each stage.44        layer_type (`str`, *optional*, defaults to `"preactivation"`):45            The layer to use, it can be either `"preactivation"` or `"bottleneck"`.46        hidden_act (`str`, *optional*, defaults to `"relu"`):47            The non-linear activation function in each block. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"`48            are supported.49        global_padding (`str`, *optional*):50            Padding strategy to use for the convolutional layers. Can be either `"valid"`, `"same"`, or `None`.51        num_groups (`int`, *optional*, defaults to 32):52            Number of groups used for the `BitGroupNormActivation` layers.53        drop_path_rate (`float`, *optional*, defaults to 0.0):54            The drop path rate for the stochastic depth.55        embedding_dynamic_padding (`bool`, *optional*, defaults to `False`):56            Whether or not to make use of dynamic padding for the embedding layer.57        output_stride (`int`, *optional*, defaults to 32):58            The output stride of the model.59        width_factor (`int`, *optional*, defaults to 1):60            The width factor for the model.61        out_features (`list[str]`, *optional*):62            If used as backbone, list of features to output. Can be any of `"stem"`, `"stage1"`, `"stage2"`, etc.63            (depending on how many stages the model has). If unset and `out_indices` is set, will default to the64            corresponding stages. If unset and `out_indices` is unset, will default to the last stage. Must be in the65            same order as defined in the `stage_names` attribute.66        out_indices (`list[int]`, *optional*):67            If used as backbone, list of indices of features to output. Can be any of 0, 1, 2, etc. (depending on how68            many stages the model has). If unset and `out_features` is set, will default to the corresponding stages.69            If unset and `out_features` is unset, will default to the last stage. Must be in the70            same order as defined in the `stage_names` attribute.71 72    Example:73    ```python74    >>> from transformers import BitConfig, BitModel75 76    >>> # Initializing a BiT bit-50 style configuration77    >>> configuration = BitConfig()78 79    >>> # Initializing a model (with random weights) from the bit-50 style configuration80    >>> model = BitModel(configuration)81 82    >>> # Accessing the model configuration83    >>> configuration = model.config84    ```85    """86 87    model_type = "bit"88    layer_types = ["preactivation", "bottleneck"]89    supported_padding = ["SAME", "VALID"]90 91    def __init__(92        self,93        num_channels=3,94        embedding_size=64,95        hidden_sizes=[256, 512, 1024, 2048],96        depths=[3, 4, 6, 3],97        layer_type="preactivation",98        hidden_act="relu",99        global_padding=None,100        num_groups=32,101        drop_path_rate=0.0,102        embedding_dynamic_padding=False,103        output_stride=32,104        width_factor=1,105        out_features=None,106        out_indices=None,107        **kwargs,108    ):109        super().__init__(**kwargs)110        if layer_type not in self.layer_types:111            raise ValueError(f"layer_type={layer_type} is not one of {','.join(self.layer_types)}")112        if global_padding is not None:113            if global_padding.upper() in self.supported_padding:114                global_padding = global_padding.upper()115            else:116                raise ValueError(f"Padding strategy {global_padding} not supported")117        self.num_channels = num_channels118        self.embedding_size = embedding_size119        self.hidden_sizes = hidden_sizes120        self.depths = depths121        self.layer_type = layer_type122        self.hidden_act = hidden_act123        self.global_padding = global_padding124        self.num_groups = num_groups125        self.drop_path_rate = drop_path_rate126        self.embedding_dynamic_padding = embedding_dynamic_padding127        self.output_stride = output_stride128        self.width_factor = width_factor129 130        self.stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, len(depths) + 1)]131        self._out_features, self._out_indices = get_aligned_output_features_output_indices(132            out_features=out_features, out_indices=out_indices, stage_names=self.stage_names133        )134 135 136__all__ = ["BitConfig"]137